What's this about?
Researchers ran ~40,000 game-like runs where people acted as the human "approver" for an AI agent's commands. When a genuinely dangerous action came up, the humans let roughly one in three of them through — evidence that asking a person to approve each agent action stops working once there's a lot of it.
What exactly is "human in the loop"?
"Human in the loop" is the common safety design where an AI can't do risky things by itself — a person must click "approve" first. The finding is about "approval fatigue": when you're bombarded with approve/deny prompts, you start rubber-stamping, so the safeguard quietly becomes theater.
What's the community saying?
HN 259 points, 194 comments.
- Confirms what many suspected: constant permission prompts train people to click "yes" without reading.
- Design fixes floated: batch/scope permissions, make dangerous actions visually distinct, default-deny the truly destructive ones.
- Pushback: is a game-run study representative of how careful people are with their real production systems and money?